search

yogeshbalaji / Generate_To_Adapt

Implementation of "Generate To Adapt: Aligning Domains using Generative Adversarial Networks"
https://arxiv.org/pdf/1704.01705.pdf
142 stars 33 forks source link

Reproduce USPS --> MNIST #4

Closed Jin-Ying closed 6 years ago

Jin-Ying commented 6 years ago

Thank you for such good work :) I have one problem when reproducing results on USPS --> MNIST: According to the network architectures and hyperparameters mentioned in the paper, the image input is 3 32 32, and the output shape of DigF2 is 48 5 5 , it's 48 5 5 = 1200 instead of 768, which cannot be the input of DigC2. Can you help me to solve this problem?

yogeshbalaji commented 6 years ago

It has to be 1200. Sorry about the typo. It will be 768 for 28x28 inputs.

Thanks.

Jin-Ying commented 6 years ago

Thanks. But after changing the first layer of DigC2 to Linear(1200,100), I still cannot reproduce the results. Could you offer me the exact input size and network architectures to reproduce the results in the paper?

OswinGuai commented 6 years ago

@yogeshbalaji After some fixes, making the code run for USPS-->MNIST, either of the results of Source-only and GTA could reach the ones shown in the paper. I am not sure whether the parameters or the architectures of netG/netD cause this. Please help. Thanks ahead!

OswinGuai commented 6 years ago

@yogeshbalaji I run that test for 200 epochs. Maybe I need more epochs?